agent-design-review

1# Agent Design Review Skill
2 
3Most agents don't fail because the model is weak — they fail because the *design* lets them loop, call the
4wrong tool, lose the thread across steps, or burn tokens with no stopping rule. This skill reviews an agent's
5architecture against the decisions that actually determine reliability, and ranks the fixes — so "it works in
6the demo but not in prod" becomes a specific list of changes. (Writing a new agent spec? Use
7[`agent-spec`](../agent-spec/SKILL.md).)
8 
9## Working from a brief
10 
11Given a sketch ("a research agent that searches, reads, and writes a report"), **deliver the full review
12anyway** — infer the likely control flow and tools, label the inference, and flag what to confirm. Never
13withhold the review for missing detail.
14 
15## Required Inputs
16 
17Ask for these only if they aren't already provided (else infer and label):
18 
19- **What the agent does** — its goal, and what a successful run produces.
20- **Control flow** — single prompt, plan-then-execute, ReAct loop, or multi-agent; and the stopping condition.
21- **Tools & actions** — what it can call, and which actions have side effects (write, send, pay).
22- **Memory & context** — what state carries across steps, and how context is kept in budget.
23- **Constraints** — latency, cost per run, and the trust boundary (untrusted input? real-world actions?).
24 
25## Output Format
26 
27### Agent Review: [agent]
28 
29**1. Summary** — will this be reliable in production? The top 3 risks and the single change that helps most.
30…